基于Multi-class SVM的车辆换道行为识别模型研究

Probe into the Multi-class SVM-based recognition model for the vehicle lane-altering behaviors

自动安全换道是车辆实现无人驾驶的关键,为精确识别行驶车辆换道状态,保证行车安全,设计了一种基于多分类支持向量机(Multi-class Support Vector Machine,Multiclass SVM)的车辆换道识别模型。从NGSIM数据集中选取美国101公路车辆轨迹数据进行分类处理,并将车辆换道过程划分为车辆跟驰阶段、车辆换道准备阶段和车辆换道执行阶段。采用网格搜索结合粒子群优化算法(Grid Search-PSO)对SVM模型中惩罚参数C和核参数g进行寻优标定,利用多分类支持向量机换道识别模型对样本数据进行训练和测试,模型测试精度达97.68%。研究表明,模型能够很好地识别车辆在换道过程中的行为状态,为车辆换道阶段的研究提供支持。

The present paper is intended to identify and propose an automatic safe lane changing model so as to accurately predict the lane changing status and ensure the traffic safety rate based on the multi-class support vector machine(short for the Multiclass SVM)so as to realize the driverless vehicle transport.For the above mentioned purpose,we have chosen No.101 road vehicle trajectory data of the USA from the NGSIM(short for the Next Generation Simulation)data set,and divided the vehicle lane changing process into the following stages,such as the vehicle lane changing preparation stage and its lane changing execution stage,including the status information changes,such as those of the speed,acceleration,and the position status-in-situ of the lane that can be extracted from the movements of the vehicles.This is because of the fact that the influence variables can represent the lane changing features we have chosen.And,in so doing,the sampling sets of the model input and the data of the variables can be normalized to reduce the dimension through the principal componential analysis.For example,the given paper has been trying to search for the penalty parameter C and the kernel parameter g in SVM model via the grid search algorithm to determine roughly the approximate interval of the optimal parameters by combining the particle swarm optimization algorithm with the optimal value calibration and accurate calculation of the model parameters so as to set up the vehicle lane changing recognition model based on the grid search and particle swarm optimization(short for the Grid Search-PSO).In addition,we have also made an effort in training the SVM categorization model in the Matlab environment with the help of the Libsvm support vector machine tool.Thus,eventually,the performance of the model has been made tested by the testing set data,with the recognition accuracy of the model being at 97.68%.In so doing,we have established the same training set and testing set data based on the unoptimized standard SVM model so as to achieve its recognition accuracy of 80.87%.The results of our research have shown that the recognition accuracy of the classification model we have proposed in the paper turns out to be 16.81%higher than that of the standard SVM one.That is to say,the classification model we have proposed can help to identify and determine the behavior status-in-situ of the vehicles during the lane-changing process and provide a firm support for the lane-changing stage quite accurately.